Using Deep Learning for Obscene Language Detection in Vietnamese Social Media

Abstract

Nowadays, a vast volume of text data is generated by Vietnamese people daily on social media platforms. Besides the enormous benefits, this situation creates many challenges. One of them concerns the fact that a tremendous amount of text contains obscene language. This kind of data negatively affects readers, especially young people. Detecting this kind of text is an important problem. In this paper, we investigate this problem using Deep Learning (DL) models such as Convolutional Neural Networks (CNN), Long-Short Term Memory (LSTM), and Bidirectional Long-Short Term Memory (BiLSTM). Besides, we combine LSTM and CNN in both sequence (sequential LSTM-CNN) and parallel (parallel LSTM-CNN) forms and sequential BiLSTM-CNN to solve this task. For word embedding phrase, we use Word2vec and PhoBERT. Experiment results show that the BiLSTM model with PhoBERT gains the best results for the obscene discrimination task, with 81.4% and 81.5% for accuracy and F1-score, respectively.